acceptodds
Under review as a conference paper at ICLR 2027

Measure Once, Steer with Many: Context Contamination and Composite Control for Persona Drift in LLMs

Abstract

Monitors and controllers of persona drift read a model's activations against a persona direction: the difference between the model's activations with and without the persona in the system prompt, extracted once before the conversation and then held fixed (the frozen direction). Re-extracted later in the dialogue, with the conversation so far as a prefix, this direction appears to rotate and shrink. We show that this movement is context contamination, not a moving persona. It follows how much new content the conversation adds: replacing the assistant's replies with another persona's barely changes it, replacing them with one repeated sentence slows it several-fold, and conversation length alone predicts it. The re-extracted direction is also the worse instrument: it does not track the judged persona within a conversation, and steering along it does worse. Steering harder along the frozen direction, however, degrades coherence and induces repetition. We therefore extract coherence and repetition as directions with the same recipe and steer along a weighted composite of all three, recovering most of the coherence and variety that pushing the persona costs. Against five published steering methods on four models, each at its best strength, some configuration of ours is at least as good on two of three measured qualities in nearly every setting. We also find that the more text accumulates, the harder the frozen direction is to express in other instructions' re-extracted directions, and that a malicious persona fades three to five times faster than a compassionate one. Our judges, blind to the method behind each reply, agree with people as closely as people agree with one another.

Primary area:
interpretability and explainable AI
Keywords:
persona drift, activation steering, persona vectors, representation engineering, interpretability, closed-loop control, multi-turn conversation, LLM evaluation, LLM-as-a-judge

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FRcvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to identify whether the current geometric/attention approaches to interpret/steer the persona drift is appropriate or not. The authors claim that the frozen direction may not match with the re-extracted direction, due to the context contamination. As more context added to the conversation, the representativity of re-extracted direction drops and becomes a bad measure for persona drift. And, they found that the persona fading/drift comes from what the models said earlier, and cannot be identified from re-extracted directions. Also, they revealed that the current geometric steering methods on frozen direction has no effect on persona drift; actually, the changes appears in the repetition instead of persona changes.

Critical strengths

This paper has the following strengths:

  • A clear and fine distinction between frozen direction and re-extracted directions, which researchers could fail to distinguish in their experiments.
  • A well-studied reasoning on why re-extracted directions are not good for identifying persona drifts, and why steering on frozen direction actually fails to lead the changes.
Critical weaknesses

This paper has the following weakness point.

  • The basis of interpretation assumes that the content itself has no or less correlation between persona itself. The assumption might not hold, because the generated content from a persona might contain useful information about the source persona. As the current reasoning heavily depends on the findings stated in Section 4 and 5, which was done on the assumption above, it should be verified whether the swapped content itself is correlated with the persona or not.
Questions for authors

See the weakness points above.

Reviewer_QYPKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether recalculating a persona direction during a conversation helps detect when an assistant goes out of character. A persona direction is the difference in average model activations for replies generated with a persona instruction and with a neutral instruction. The direction measured in separate runs before the conversation tracked judges’ ratings better than one recalculated using the conversation so far. The authors also combine directions for persona, coherence, and repetition to reduce the side effects of steering for persona alone.

Critical strengths
  • Comparing replies of the same length clarifies why a shrinking persona direction can be a misleading sign of drift. The recalculated direction retained 8–9% of its initial length with either persona’s replies, compared with 48% with repeated text.

  • It is encouraging that the steering settings carried over to new conversations without retuning. Combined steering matched or outperformed persona-only steering on coherence in 56 of 72 held-out conversations and on variety in 60.

Critical weaknesses
  • It remains unclear how much combining steering directions adds over simply reducing the strength of persona-only steering. On held-out conversations, combined steering improved coherence but scored 1.9 points lower on persona adherence on average. At comparable adherence, Figure 5 suggests only a tentative gain in reply variety.

  • Prior work used fixed directions to monitor character traits and persona drift (Chen et al., 2025, Lu et al., 2026). Here, the monitor and judge scores may both decline over time, so their 0.48 correlation leaves open whether the monitor detects individual off-character replies.

Questions for authors
  • Within each persona and user stance, how well do the fixed and recalculated readings predict judges’ ratings of individual replies after accounting for turn number?

  • At comparable persona adherence, how large and consistent are the differences in coherence and repetition between combined steering and weaker persona-only steering on the same held-out conversations?

3 comments
AuthorsOpenReview ↗

Fixed and recalculated readings against the judges' ratings of individual replies, and combined vs weaker persona-only steering at comparable persona adherence

Thank you for the two precise questions. We ran new analyses for both, now in the revised paper, and following are the details.

1. Fixed vs recalculated readings within each persona and user stance (Q1, W2)

A reading is the dot product of the model's activation just before its reply with a persona direction's unit vector. The fixed reading uses the frozen direction v0v_0, extracted once before the conversation. Two recalculated readings use vtv_t, re-extracted at turn tt with the conversation so far as a prefix: the whole re-extracted reading uses vtv_t itself, and the contaminated reading uses only the part of vtv_t perpendicular to v0v_0 -- the contaminated component, which is all that re-extraction adds.

To show how well each reading predicts the judged persona adherence of individual replies, Figures S7 and S8 (Appendix L) plot the 18 unsteered conversations of Section 4.2 at every turn from 1 to 40, three per persona and user stance. At each turn we average each reading, and the judged persona adherence (gpt-oss-120b), over the three conversations. Over the 40 turns, the Spearman correlation of each averaged reading with the averaged judged persona adherence is:

PersonaUser stanceFixedWhole re-extractedContaminated
CompassionateSupportive0.370.200.12
CompassionateNeutral0.360.330.25
CompassionateOpposed0.760.680.58
MaliciousSupportive0.630.550.34
MaliciousNeutral0.700.560.36
MaliciousOpposed0.630.530.43

So in every row, the fixed reading follows the judged persona adherence most closely, the whole re-extracted one less closely and the contaminated one least closely. Once the malicious persona has faded, the judged persona adherence and the fixed reading stay low, but the contaminated reading climbs back over the second half of the conversation and pulls the whole re-extracted reading up with it.

To test how strongly each direction moves individual replies toward the persona, we steered along each (Figure S9). In each conversation, at turns 5, 10, ..., 40 and with two seeds (288 cases), we regenerated the reply without a push and with an equal push along each direction. At each case's best of three push sizes, judged persona adherence rises above the unsteered reply by 58 points on average along the frozen direction, 48 along the whole re-extracted direction, 27 along its contaminated component and 5 along a random one (gpt-oss-120b, Table S18(b)). The order holds under Claude Sonnet 5 and GPT-5-mini too, and on the same cases the frozen direction beats the whole re-extracted one by 10, 8 and 6 points under the three judges, with every 95% interval above zero.

2. Combined vs weaker persona-only steering at comparable persona adherence (Q2, W1)

It is true that combined steering -- our composite -- gains coherence at a small cost in persona adherence (1.9 points in Table S20), so it is a trade-off. On all 72 held-out conversations, we reran the tuned composite and ran persona-only steering at its tuned strength and at two other strengths of its tuning grid, almost always the two just below the tuned one. From these three runs we interpolated persona-only steering's coherence and variety at the composite's persona adherence. Variety is 100 × (1 − repetition).

Under gpt-oss-120b, the composite's persona adherence lies within the range of the three persona-only runs in 42 conversations, and above all three, even the tuned one, in 23 of the other 30. In these 42, persona-only steering at the same persona adherence is no more coherent or varied than the composite: the composite's coherence differs by 0.0 points (95% interval −1.5 to +1.5) and its variety by +0.7 (−2.1 to +3.3), and both intervals still include zero if any one of the four models is removed. Claude Sonnet 5 also finds no difference, and GPT-5-mini even finds the composite 3.5 points more varied (+0.9 to +6.2), with no difference in coherence.

What the composite adds is a better balance at its tuned weights and, in its adaptive form, less damage when pushed hard:

  • With both tuned on the same tuning conversation, the composite gives up 0.9 points of persona adherence in these reruns (−2.1 to +4.3) but gains 6.8 points of coherence and 11.4 of variety (gpt-oss-120b) -- more coherent in 57 and more varied in 58 of the 72 conversations.
  • In its adaptive form it needs no weight search and steers exactly along the frozen persona direction until coherence or repetition gets worse.
  • Past persona-only steering's peak in persona adherence, on the tuning conversations, the adaptive composite keeps 4.5 more coherence points than persona-only steering at the same strength (gpt-oss-120b; +2.4 to +6.7, more in 22 of 28 settings; the other judges agree), with 3.9 points higher persona adherence.

Thank you again -- both questions made the paper more precise. We hope this resolves your concerns, and we would be grateful to get further feedback from you!

Thank you again for your careful review and your two precise questions. We wanted to gently follow up while the discussion period is still open. We ran new analyses for both questions, and they are now in the revised paper. We would be grateful to know if this resolves your concerns, and if so, we would really appreciate it if you would consider raising your score.

Dear Reviewer QYPK,

As the discussion period is nearing its end, we sincerely hope that our responses have adequately addressed your concerns. We greatly appreciate your time and thoughtful feedback, and we remain available to clarify any outstanding questions or concerns.

Thank you again for your valuable input.

Reviewer_BWbBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This work examines how persona-related activation directions behave, as an LLM conversation grows. Firstly, the frozen steering direction extracted before conversation is compared to the direction re-extracted across subsequent conversation turns, and it is shown that re-extracted directions shrink and rotate as the conversation goes on. This is attributed to context contamination rather than persona representation causes. Steering along a single persona direction eventually leads to degradation / repetition, and the authors combine three activation directions (persona adherence, coherence, repetition) into a controller.

Critical strengths
  1. This work is an interesting piece of mechanistic exploration, and can help researchers better understand how persona is quantified, how it can be steered, and how context contamination can occur.
Critical weaknesses
  1. The experiments / results cannot conclusively establish that the underlying persona itself remains stationary. The re-extracted directions throughout the conversation, could genuinely be sensitive to additional context. This means that the representation of a persona itself could have changed in some aspects or be contaminated, and the re-extracted directions themselves could also be contaminated. This is a serious problem in the experimental set-up. The paper makes a lot of claims that are grander than the results can support.

  2. Many issues with the experimental set-up and results, including evaluation-time tuning (Appendix M), unequal baseline search budgets (Appendix N), selective layer analysis (Appendix B). None of them are fatal flaws by themselves; the biggest issue is that the paper's mechanistic claims exceed what the results can concretely show.

  • less relevant, but the writing in the main body of the paper should be polished for human readability. Also, I think the 36-page Appendix is an absolute overkill - a good paper is one that can succinctly convey information to a human reader, and present only the important aspects of it. The captions for the tables / figures are also incredibly long and difficult to read, each is almost an entire paragraph in itself.
Questions for authors
  1. Mechanistic validity. In Sections 4-6 (and I suppose Appendix D), the experiments show that re-extracted directions are sensitive to accumulated text. However, this does not rule out genuine changes in persona representations. This makes the entire set-up of your paper precarious, because how can you separate context contamination from actual representation drift? What justifies your conclusion that the persona direction itself is not moving?

  2. The primary layers are selected where preliminary results showed the greatest drift, and this is acknowledged by the authors. How robust are your findings to independently / randomly selected layers and probe sets?

3 comments

Thank you again for your thoughtful review. We wanted to gently follow up before the discussion period ends. We have revised the paper based on your feedback, stating our claims more precisely and adding results at every layer. We would be grateful to know if this addresses your concerns, and if it does, we hope you might consider raising your score.

Thanks for running the additional experiments! In particular, I think testing separating context contamination from representation drift, and testing out the layers and probe sets was important. (That said, there really has to be a better way to incorporate them than adding it to your already massive appendix.)

At the core of it though, I am still concerned with how you distinguish representation validity (which this paper does!) vs. intervention efficacy. For one, a direction can still be casually effective for steering outputs, even if it doesn't track the representation naturally used by the model. While you can recover persona adherence through steering, it doesn't establish that the frozen direction continues to represent the model's unsteered state. The re-injection is also not a clean control, and it restores the norm but not the original direction. I do think there's an explanation for the direction change, which needs to be explored deeper.

The latest PDF has a slight improvement for human readability, but there are still many long, wordy paragraphs and the captions still comprise of multiple sentences. The supplementary materials is still a whooping 36 pages long - it really makes following the work and referencing results very tiresome. I will raise the presentation score for your efforts in the rebuttal, and I think it would be nice to break some of the super long paragraphs up in the next revision. Thank you for your time and efforts on the rebuttal.

AuthorsOpenReview ↗

Representation validity vs. intervention efficacy; why we re-injected the persona instruction; investigating the direction change

Thank you for raising the presentation score!

1. Representation validity vs. intervention efficacy

It is true that steering on its own only shows intervention efficacy, not representation validity. For our paper, this comes down to what we actually need a persona representation for. We only need it for two things -- to monitor how much of the persona the model expresses in its replies, and to steer the model's activations to improve persona adherence. These are the measurement and intervention notions of a linear representation (Park et al., 2024).

  • Monitoring is tested on the model's unsteered state: Judged persona adherence is tracked most closely by the projection onto the frozen direction (Spearman 0.48 under gpt-oss-120b), less closely by the projection onto the whole re-extracted direction (0.41) and least closely by the projection onto its contaminated component, the part perpendicular to the frozen one (0.35; Table 2, Figure S2).
  • Steering is tested separately: With equal pushes, judged persona adherence rises above the unsteered reply by 58 points along the frozen direction, 48 along the whole re-extracted one, 27 along the contaminated component and 5 along a random one (gpt-oss-120b; Table S18(b)).

The frozen direction does both tasks, and comes first in both, and thus it satisfies all that's needed from a persona representation for our paper's claims. You are correct that there might be some other task or purpose -- for example identifying the representation the model naturally uses to produce the persona -- which might not be correctly showcased by the frozen direction. For those, one would have to find a representation that serves them too, as we noted for future work in our earlier reply. In the second revision of our paper, we now state this scope in the abstract, Section 4.2 and the Limitations.

2. Why we re-injected the persona instruction, and the direction change

Our first reply did not explain why we re-injected the persona instruction. The re-injection was not meant to bring back the original persona direction. We used it to check whether the re-extracted direction shrinks and turns only because the instruction gets far from the probe question (point of re-extraction) as the conversation grows. If that were all, putting the instruction next to the probe question would undo both effects. It does not -- the length comes back (91% of the frozen length at turn 40, against 11% with the instruction in the system prompt), but the direction stays turned away from the frozen one (cosine 0.72). So re-injecting the instruction does not remove the contamination: the conversation's content is still in the prefix and gets mixed in. This is something that we had expected.

We agree the direction change deserves a deeper explanation. So far we know the re-extracted direction follows the conversation's content, not how much of the persona the model expresses:

  • it follows new content (diversity in content): at turn 40, its cosine with the frozen direction is 0.37-0.43 with natural replies, but 0.52-0.59 with repeated or filler ones (malicious persona; Table S6);
  • if we fit a curve in conversation length, blind to the instruction, it predicts how far the direction has turned (Appendix E);
  • steering along the frozen direction keeps the persona expressed (judged persona adherence 86 against 11 unsteered), but the re-extracted direction turns away just as much: cosine 0.46 at turn 40 (0.48 unsteered; Qwen2.5-14B).
  • re-weighting attention toward the system prompt also restores the persona (87), but this causes the re-extracted direction to stay closer to the frozen one at turn 40, on all three models tested (Table S24).

So the re-extracted direction gets affected by the mixing of the persona instruction with the conversation's content: diversity in content moves it, and re-weighting attention to the persona instruction can limit its movement. Where in the model this mixing happens is our next research question which we have noted as future work.

3. Presentation

Following your first point, main paper's Table 2 and Figure 2 now show monitoring on unsteered conversations next to steering. Every caption is now one sentence, and the long paragraphs are split and their numbers are tabularized in the main paper. For the camera-ready version, we will move the key panels of Figures S2, S7, S8 and S9 into the main paper and shorten the supplementary material.

If this resolves your concerns, we would be grateful if you would consider raising your rating. Your support means a lot to us.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of changes in the revised paper

We have submitted our revised paper -- main paper and supplementary material. Following are the notable changes between the original and the revised PDFs. Section, figure and table numbers refer to the revised PDFs.

New analyses

  • We split the re-extracted direction into its component along the frozen direction and its component perpendicular to the frozen direction, which we call its contaminated component (Section 3, Appendix A). This helped us explicitly showcase the effects of context contamination on the performance of the re-extracted persona direction.
  • Section 4.2 and Appendix D: re-extracting with the persona instruction next to the probe question removes the shrinkage but not the rotation of the re-extracted persona direction.
  • Appendix B: the results at every layer and with five question sets; Section 3 points to them.
  • Appendix L: the readings at every turn from 1 to 40 (Figures S7 and S8), and steering along the frozen direction, the whole re-extracted direction, its contaminated component and a random direction (Table S18(b), Figure S9; Section 4.2 gives the results); paired on the same cases, the frozen direction's gain exceeds the whole re-extracted direction's under every judge.
  • Section 8 and Appendix M: the composite against weaker persona-only steering at comparable persona adherence, and the composite's properties, under all three judges.
  • Appendix N: the comparison with published methods when each side gets the same number of configurations.

Claims stated more precisely

  • Introduction and Section 4: the re-extracted direction's movement follows the context of the conversation. It is not evidence that the persona representation changed.
  • Section 8: the held-out results come first, and results with each setting's own best configuration are labelled as upper bounds. The composite repairs much of what persona-only steering costs, at a small cost in persona adherence. The comparison with published methods is described as a trade-off. Figure 5's claim is narrowed: pooled over the four models, the composite reaches combinations the re-extracted controller does not, and against the frozen controller this is a tendency.

Presentation

  • Captions are cut to their main point, and the wording of the main text and the supplementary is simplified and made more precise throughout.
  • Figures S6, S10 and S12 are redrawn so that every label is legible at print size.
  • The extended related work (Appendix R) is removed.
  • We are continually working to enhance the readability and conciseness of our paper, and these improvements will be reflected in future revisions.
AuthorsOpenReview ↗

Summary of changes in second revision of our paper

We have uploaded a second revision of the main paper and the supplementary material. It states more precisely what the persona direction is used for and makes the presentation easier to follow; no result changed. Section, figure and table numbers below refer to revision 2.

  • Scope of the "persona representation validity" claim: the abstract, Section 4.2, the Conclusion and the Limitations now say that the frozen direction is a valid representation of the persona for monitoring persona adherence and steering to regain persona adherence, and as to whether the model itself relies on it to produce the persona remains open.
  • Figure 2 is new. For the malicious persona, it shows (a) the activation projections onto each persona direction and judged persona adherence on the unsteered conversations (monitoring; from Figure S8) and (b) how a push along each persona direction raises judged persona adherence by turn (steering; from Figure S9). Table 2 (new) gives the numbers for both.
  • Section 4.2 and Appendix D now state why we re-injected the persona instruction: to test whether the movement showcased by the re-extracted persona direction comes only from the instruction's growing distance from the point of re-extraction (probe point). Placed next to the probe point, the instruction restores the length but not the direction.
  • New tables replace dense paragraphs: the conversation sets (Table 1), the three directions as monitors and as steering levers (Table 2), the composite against persona-only steering (Table 3) and the published (baseline) methods (Table 4).
  • Revision 1's Figure 2 is now Figure 3. Figure 4 now shows stacked bars; its per-setting version is Figure S13. The caption of Figure 5 is now one sentence. The per-setting radar (revision 1's Figure 3) and the contraction table (revision 1's Table 1) moved to the supplementary material (Figure S12 and Table S10). Every figure and table number of the supplementary material that our replies cite here, is unchanged.
  • Presentation: a shorter abstract; every caption, in both documents, is one sentence that opens with its takeaway (in the supplementary material, legend details are in a small-print note under each figure or table); long paragraphs are split; shorter sentences throughout, with displayed equations for the split of the re-extracted reading (Section 3), the controller (Section 7) and the composite (Section 8).
  • For the camera-ready version, we will move the key panels of Figures S2, S7, S8 and S9 into the main text and shorten the supplementary material.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.